Hybrid Feature Model Based Knee Vibroarthrographic Signal Classification with Signal Level Enhancement Using Morphological Filter

When the human knee is in motion, joint vibration signals are detected and recorded. These vibration signals termed as a Vibroarthrographic (VAG) signals are helpful in assessing the knee joint pathological condition. So, by employing feature extraction and machine learning algorithms, VAG signals can be classified into normal and abnormal, which will assist in non-invasive investigation for medical practitioners. In this paper, we propose a unique morphological denoising filters on VAG signal to improve signal to noise ratio (SNR). Along with gold standard statistical features, features such as RMS values of opening, closing and average morphological filters are computed. Fifteen statistical features are reduced to four using Fischer score algorithm. In this hybrid model of feature combination, four spectral features which are spectral slope (variance), spectral flux (mean), spectral flux (variance) and spectral mean (variance) are also a part of final feature set. Pattern classification performance is tested and compared using Random Forest (RF), Gaussian Naïve Bayes (GNB) and Support vector Machine (SVM-RBF) classifiers. Random forest model performed best offering classification accuracy of 93.07%, sensitivity 90% and specificity of 87.88%.